[国内学会] Investigation of WaveNet for Text-to-Speech Synthesis

情報処理学会 第118回音楽情報科学・第120回音声言語情報処理合同研究発表会

WaveNet is a type of neural network that can be used to model speech waveforms. It has been used in text-to-speech synthesis systems to convert acoustic or linguistic features into waveforms. Despite the description in recent literatures and open-source implementation, the mechanism of WaveNet is still somewhat obscure. This work explains the authors’ WaveNet implementation. It also introduces a one-best generation method that could be an alternative to the random-sampling-based generation method. Based on the implementation, this work shows observations inside the network. Interesting findings include the manifold of quantized waveforms learned by WaveNet and the gradually decreased data variance in WaveNet blocks. These results may be helpful for further investigation on WaveNet.